⚡ electrical engineering

From Manual Observation to Automated Monitoring: Space Allowance Effects on Play Behaviour in Group-Housed Dairy Calves

This study demonstrates that play behavior in group-housed dairy calves peaks at a space allowance of 8–10 m² per calf and can be effectively monitored at scale using a highly accurate automated computer vision pipeline, offering a practical tool for optimizing calf welfare and economic feasibility.

Original authors: Haiyu Yang, Heidi Lesscher, Enhong Liu, Miel Hostens

Published 2026-02-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haiyu Yang, Heidi Lesscher, Enhong Liu, Miel Hostens

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of young calves as energetic toddlers in a daycare. Just like children, they have a natural urge to run, jump, wrestle, and play. In the world of dairy farming, this "playtime" is a secret code: if the calves are playing, they are happy and healthy. If they are bored or sick, they stop playing.

For a long time, farmers and scientists have argued about how much floor space each calf needs to be happy. The current rules say a tiny amount of space is enough, but animal welfare experts suggest they need a lot more. The problem? No one really knew what the "sweet spot" was, and watching calves all day to see if they were playing is like trying to count every grain of sand on a beach by hand—it takes forever and is easy to get wrong.

This study tackled two big questions: How much space do calves actually need to play? and Can we use computers to watch them for us?

1. The "Goldilocks" Zone for Space

The researchers went to 14 different farms in the Netherlands and looked at calves living in pens of all different sizes, from very cramped (about the size of a small bathroom) to very spacious (like a large living room).

They found that the relationship between space and play isn't a straight line where "more space = more play." Instead, it's more like a rollercoaster:

Too Small: When calves were very crowded, they didn't play much.
Just Right: The happiest, most playful calves were in pens with 8 to 10 square meters of space per calf. In these pens, calves played for about 16 minutes a day.
Too Big (or just weird): Surprisingly, when the pens got even bigger (12–14 square meters), playtime actually dropped. It went back up slightly in the biggest pens, but it never got better than the "Goldilocks" zone.

The Analogy: Think of it like a dance floor. If the room is too small, everyone bumps into each other and can't dance. If the room is huge and empty, people feel awkward and stand around. But if the room is just the right size, everyone feels comfortable enough to let loose and dance. For calves, that perfect dance floor is 8–10 square meters.

2. The "Robot Watcher" (Computer Vision)

Watching 64 calves for 17 hours a day is a massive job. Doing it by hand is slow and prone to human error. So, the team built a computer brain to do the watching.

They used a "pipeline" of smart tools (like a team of specialized robots):

The Eyes (YOLOv12): This tool spots the calves in the video, like a security guard spotting people in a crowd.
The Tracker (SAM2): Once it spots a calf, this tool puts a digital "halo" around it and follows it as it moves, making sure it knows which calf is which, even if they run behind a fence.
The Brain (DinoV2): This is the super-smart part. It looks at the movement and decides, "Is this a play move (like a kick or a jump) or just a normal move (like walking to eat)?"

The Result: The computer was incredibly accurate. It got the right answer 97.6% of the time. Even better, it caught 99.4% of the actual play moments. It was so good that it could learn from a small amount of human-made notes and then apply that knowledge to watch thousands of hours of video automatically.

3. What They Found About Play

Social is King: The most common type of play was social—calves pushing each other or chasing. They loved to play together.
The "Burst" Nature: Calves didn't play for hours straight. They had short bursts of energy, like a toddler having a "zoomies" moment, running and jumping for a few seconds, then stopping.
Health Check: The study confirmed that play is a great health barometer. Sick or stressed calves simply didn't play.

The Takeaway

This paper doesn't just tell us that calves need space; it gives us a specific number: 8 to 10 square meters per calf seems to be the practical target for maximizing their happiness without breaking the bank.

It also proves that we don't need armies of people to watch animals anymore. With the right cameras and smart software, we can turn a small, expensive research project into a system that can monitor animal welfare continuously, 24/7, on farms of any size.

In short: Give calves a medium-sized room (not too small, not too huge), and they will dance. And now, we have a robot that can watch them dance and tell us if they are having a good time.

Technical Summary: From Manual Observation to Automated Monitoring of Play Behaviour in Group-Housed Dairy Calves

Problem Statement

Animal welfare in modern dairy production is increasingly defined by the capacity to facilitate positive emotional states and species-typical behaviors, such as play. While play serves as a robust indicator of positive welfare, its relationship with space allowance under commercial conditions remains poorly characterized, particularly at intermediate-to-high densities (6–20 m² per calf). Current European Union legislation mandates a minimum of 1.5–1.8 m², whereas recent EFSA recommendations suggest 20 m² for full locomotor play expression; however, empirical data bridging this gap is scarce. Furthermore, systematic monitoring of play on commercial farms is hindered by the labor intensity, observer bias, and lack of scalability inherent in manual video analysis. Existing sensor-based approaches (e.g., accelerometers) lack the specificity to distinguish play from other activities and fail to provide spatial context.

Methodology

Study Design and Data Collection

The study investigated 60 group-housed dairy calves across 14 commercial farms in the Netherlands. Space allowances ranged from 2.66 to 17.98 m² per calf. Data collection occurred over one-week periods per farm (June 2024–May 2025), involving continuous video recording (06:00–23:00) via ceiling- or side-mounted Bascom cameras.

Ground Truth Creation: Seven trained observers manually annotated video footage using Observer XT software. A detailed ethogram categorized play into locomotor, social, object, and straw-related play. Observations were recorded with 0.1-second resolution, focusing on the second day of each visit to minimize observer effects.
Covariates: Data included calf demographics (age, weight, breed), health status (categorized by symptom severity), milk intake, group size, and bedding quality. Calves with severe illness were excluded.

Computer Vision Pipeline Development

To enable scalable monitoring, a pipeline integrating three state-of-the-art foundation models was developed:

Object Detection: YOLOv12 (pre-trained on COCO) detected calf bounding boxes.
Segmentation and Tracking: SAM2 generated segmentation masks and tracked individuals across frames using video memory.
Feature Extraction: DinoV2 (a vision transformer pre-trained via self-supervised learning on 142 million images) extracted 1024-dimensional visual embeddings from cropped calf images.
Metadata Integration: Manual annotations were temporally aligned with video frames using Optical Character Recognition (OCR) to extract timestamps from on-screen overlays.
Classification: A Multi-Layer Perceptron (MLP) classifier was trained on the DinoV2 embeddings to predict behavior classes (Active Playing, Non-Active Playing, Not Playing). The dataset was stratified to address class imbalance, resulting in a balanced training set of ~22,800 samples.

Statistical Analysis

Linear mixed models were employed to analyze the relationship between space allowance and play duration (expressed as percentage of observation period, %OP), with farm included as a random effect. Multivariable models controlled for age, health, milk intake, group size, and bedding quality.

Key Results

Space Allowance and Play Behaviour

Overall Play: Calves spent an average of 1.0% of the observation period playing (approx. 10 minutes per 17-hour day). Social play constituted the majority (55% of total play time), followed by locomotor (26%), object (15%), and straw play (2%).
Non-Linear Relationship: A significant, non-linear relationship was found between space allowance and total play duration ( $P < 0.01$ $P < 0.01$ ).
- Peak Performance: The highest play levels occurred at 8–10 m² per calf (1.59% OP, ~16 minutes/day).
- Low Points: Play was lowest at 6–8 m² and 12–14 m² (<0.6% OP).
- Plateau: At 16–18 m², play levels (1.29% OP) did not significantly exceed those at 8–10 m², suggesting a plateau effect rather than a linear increase.
Subtype Specificity: Space allowance significantly predicted social play duration ( $P < 0.05$ ), with peaks at 8–10 m² and 16–18 m². No significant effect was found for object play, and the effect on locomotor play was less pronounced than for total play.

Automated Monitoring Performance

Classification Accuracy: The MLP classifier achieved 97.6% overall accuracy on the held-out test set.
Recall for Active Play: The model demonstrated 99.4% recall for active play detection, indicating a very low rate of false negatives, which is critical for welfare monitoring.
Precision: Precision for active play was 0.96.
Pipeline Efficiency: Processing 90,000 frames (5 hours of video) required approximately 35 hours of compute time on a Databricks cluster (7:1 processing-to-video ratio). The system successfully processed 108 hours of footage from 6 farms for training.

Technical Challenges Identified

The study highlighted specific deployment challenges in commercial environments:

Camera Angles: Overhead cameras provided superior tracking stability (100% frame success) but suffered from low initial detection rates (25%) due to domain shift from eye-level training data. Side-mounted cameras had higher detection rates but lower tracking stability due to occlusions.
White Object Hallucination: SAM2 exhibited "white object hallucination," where white calves on white straw bedding caused segmentation masks to leak into the background, leading to tracking failures.
Temporal Alignment: OCR success rates for timestamp extraction varied widely (3% to 99.8%) depending on font contrast and rendering quality, affecting the alignment of ground truth with video frames.

Key Contributions and Significance

1. Empirical Evidence for Space Standards:
The study provides the first comprehensive characterization of the space-play relationship across a broad gradient (2.7–18 m²) under commercial conditions. It challenges the assumption that more space linearly equates to more play, identifying 8–10 m² per calf as a practical target that maximizes play behavior. This finding suggests that current EU minimums (1.5–1.8 m²) are inadequate, while the EFSA recommendation of 20 m² may exceed the threshold where space is the primary limiting factor for play.

2. Scalable Automated Monitoring:
The research demonstrates that computer vision pipelines utilizing foundation models (YOLOv12, SAM2, DinoV2) can scale small manual annotation projects (108 hours) into continuous, objective welfare assessment systems. The system achieves high fidelity (97.6% accuracy) with minimal task-specific training, overcoming the scalability limitations of manual observation.

3. Practical Implementation Guidelines:
The paper offers evidence-based recommendations for deploying automated monitoring, emphasizing the trade-offs between camera mounting angles (overhead for stability vs. side-mounted for detection) and the necessity of high-contrast timestamp overlays for temporal alignment.

4. Modest Claims on Causality:
The authors explicitly acknowledge the observational nature of the study, noting that while statistical models controlled for measured confounders, unmeasured farm-level factors (e.g., management style, enrichment) may influence the results. The study does not claim to prove causality but provides high ecological validity data to inform welfare standards and future randomized trials.

In summary, this work bridges animal behavior science and technological innovation, offering a validated method to monitor positive welfare indicators at scale and providing data-driven guidance for optimizing housing space to enhance calf well-being.